Does your LLM truly unlearn? An embarrassingly simple approach to recover unlearned knowledge

Large language models (LLMs) have shown remarkable proficiency in generating text, benefiting from extensive training on vast textual corpora. However, LLMs may also acquire unwanted behaviors from the diverse and sensitive nature of their training data, which can include copyrighted and private content. Machine unlearning has been introduced as a viable solution to remove the influence of such problematic content without the need for costly and time-consuming retraining. This process aims to erase specific knowledge from LLMs while preserving as much model utility as possible. Despite the effectiveness of current unlearning methods, little attention has been given to whether existing unlearning methods for LLMs truly achieve forgetting or merely hide the knowledge, which current unlearning benchmarks fail to detect. This paper reveals that applying quantization to models that have undergone unlearning can restore the "forgotten" information. To thoroughly evaluate this phenomenon, we conduct comprehensive experiments using various quantization techniques across multiple precision levels. We find that for unlearning methods with utility constraints, the unlearned model retains an average of 21\% of the intended forgotten knowledge in full precision, which significantly increases to 83\% after 4-bit quantization. Based on our empirical findings, we provide a theoretical explanation for the observed phenomenon and propose a quantization-robust unlearning strategy to mitigate this intricate issue...

Enhance Graph Alignment for Large Language Models

Graph-structured data is prevalent in the real world. Recently, due to the powerful emergent capabilities, Large Language Models (LLMs) have shown promising performance in modeling graphs. The key to effectively applying LLMs on graphs is converting graph data into a format LLMs can comprehend. Graph-to-token approaches are popular in enabling LLMs to process graph information. They transform graphs into sequences of tokens and align them with text tokens through instruction tuning, where self-supervised instruction tuning helps LLMs acquire general knowledge about graphs, and supervised fine-tuning specializes LLMs for the downstream tasks on graphs. Despite their initial success, we find that existing methods have a misalignment between self-supervised tasks and supervised downstream tasks, resulting in negative transfer from self-supervised fine-tuning to downstream tasks. To address these issues, we propose Graph Alignment Large Language Models (GALLM) to benefit from aligned task templates. In the self-supervised tuning stage, we introduce a novel text matching task using templates aligned with downstream tasks. In the task-specific tuning stage, we propose two category prompt methods that learn supervision information from additional explanation with further aligned templates. Experimental evaluations on four datasets demonstrate substantial improvements in supervised learning, multi-dataset generalizability, and particularly in zero-shot capability, highlighting the model's potential as a graph foundation model.

HC-GST: Heterophily-aware Distribution Consistency based Graph Self-training

Graph self-training (GST), which selects and assigns pseudo-labels to unlabeled nodes, is popular for tackling label sparsity in graphs. However, recent study on homophily graphs show that GST methods could introduce and amplify distribution shift between training and test nodes as they tend to assign pseudo-labels to nodes they are good at. As GNNs typically perform better on homophilic nodes, there could be potential shifts towards homophilic pseudo-nodes, which is underexplored. Our preliminary experiments on heterophilic graphs verify that these methods can cause shifts in homophily ratio distributions, leading to \textit{training bias} that improves performance on homophilic nodes while degrading it on heterophilic ones. Therefore, we study a novel problem of reducing homophily ratio distribution shifts during self-training on heterophilic graphs. A key challenge is the accurate calculation of homophily ratios and their distributions without extensive labeled data. To tackle them, we propose a novel Heterophily-aware Distribution Consistency-based Graph Self-Training (HC-GST) framework, which estimates homophily ratios using soft labels and optimizes a selection vector to align pseudo-nodes with the global homophily ratio distribution. Extensive experiments on both homophilic and heterophilic graphs show that HC-GST effectively reduces training bias and enhances self-training performance.

InfuserKI: Enhancing Large Language Models with Knowledge Graphs via Infuser-Guided Knowledge Integration

Though Large Language Models (LLMs) have shown remarkable open-generation capabilities across diverse domains, they struggle with knowledge-intensive tasks. To alleviate this issue, knowledge integration methods have been proposed to enhance LLMs with domain-specific knowledge graphs using external modules. However, they suffer from data inefficiency as they require both known and unknown knowledge for fine-tuning. Thus, we study a novel problem of integrating unknown knowledge into LLMs efficiently without unnecessary overlap of known knowledge. Injecting new knowledge poses the risk of forgetting previously acquired knowledge. To tackle this, we propose a novel Infuser-Guided Knowledge Integration (InfuserKI) framework that utilizes transformer internal states to determine whether to enhance the original LLM output with additional information, thereby effectively mitigating knowledge forgetting. Evaluations on the UMLS-2.5k and MetaQA domain knowledge graphs demonstrate that InfuserKI can effectively acquire new knowledge and outperform state-of-the-art baselines by 9% and 6%, respectively, in reducing knowledge forgetting.

Distribution Consistency based Self-Training for Graph Neural Networks with Sparse Labels

Few-shot node classification poses a significant challenge for Graph Neural Networks (GNNs) due to insufficient supervision and potential distribution shifts between labeled and unlabeled nodes. Self-training has emerged as a widely popular framework to leverage the abundance of unlabeled data, which expands the training set by assigning pseudo-labels to selected unlabeled nodes. Efforts have been made to develop various selection strategies based on confidence, information gain, etc. However, none of these methods takes into account the distribution shift between the training and testing node sets. The pseudo-labeling step may amplify this shift and even introduce new ones, hindering the effectiveness of self-training. Therefore, in this work, we explore the potential of explicitly bridging the distribution shift between the expanded training set and test set during self-training. To this end, we propose a novel Distribution-Consistent Graph Self-Training (DC-GST) framework to identify pseudo-labeled nodes that are both informative and capable of redeeming the distribution discrepancy and formulate it as a differentiable optimization task. A distribution-shift-aware edge predictor is further adopted to augment the graph and increase the model's generalizability in assigning pseudo labels. We evaluate our proposed method on four publicly available benchmark datasets and extensive experiments demonstrate that our framework consistently outperforms state-of-the-art baselines.